Abstract: The diversity of land cover types in complex urban environments increases the difficulty of high-precision extraction of urban vegetation information from remote sensing images.However, the proposal of deep learning theory and the application of its methods have brought a new perspective to the extraction of vegetation information from remote sensing images.The paper addresses the issue of insufficient utilization of spatial contextual information in the traditional U-Net neural network model and improves it by incorporating attention mechanisms to reduce interference from complex backgrounds.Additionally, the Atrous Spatial Pyramid Pooling（ASPP） structure is added to better integrate contextual information, and residual connections are introduced in the convolutional modules to effectively alleviate the problems of gradient vanishing and information loss caused by multiple convolutions.The paper conducts experiments and accuracy evaluations based on the ISPRS（Potsdam） dataset.The results show that the improved U-Net model improves the accuracy, intersection ratio, F1 score and Kappa coefficient compared with the traditional U-Net model, The improved U-Net model can achieve higher precision in urban vegetation extraction from high-resolution remote sensing images.